ACCELERATION OF FACE LOCALIZATION ALGORITHM FROM OPENCV LIBRARY ON OPERATING SYSTEM ANDROID

V diplomskem delu smo se ukvarjali s pohitritvijo postopkov za lokaliziranje obrazov. Pregledno smo predstavili obstoječe algoritme za zaznavanje obrazov iz knjižnice OpenCV. V praktičnem delu naloge smo spremenili osnovni algoritem za zaznavanje obrazov iz knjižnice OpenCV, prirejen za mobilne naprave z operacijskim sistemom Android, da algoritem za ceno manjše natančnosti deluje hitreje. Osnovni algoritem je splošen in se lahko uporablja na vseh napravah, ki podpirajo izvajanje kode C++. Zaradi želje po splošnosti se algoritem lahko izvaja na različnih napravah, a delovanje tega algoritma na nobeni ni posebej hitro. V diplomskem delu smo ta algoritem spremenili iz splošno namenskega v specifičnega tako, da smo mu odvzeli določene manj pomembne funkcionalnosti, priredili kodo za naprave Android, ob tem pa smo optimizirali še hitrost izvajanja, vse na račun malenkostnega zmanjšanja natančnosti lokaliziranja obrazov. Eksperimentalni rezultati so pokazali, da s prilagajanjem algoritma ciljni napravi takšen algoritem deluje bistveno hitreje (do 13 ms hitreje kot izvirni algoritem pri lokaliziranju obrazov na enem posnetku) na račun zmanjšanja uspešnosti lokaliziranja obrazov za 18 %.In the thesis we were working on accelerating algorithms for face localization. We reviewed existing algorithms from the open source computer vision library OpenCV. We changed the basic face detection algorithm from the OpenCV library, adapted for Android mobile devices, in such a way that the algorithm runs faster, for the price of lower accuracy. The basic algorithm is general and can be used on any device that supports C++ code execution. Because of that requirement of generality, the algorithm is indeed suitable for many devices, but its speed is lacking on all of them. In this thesis, we changed this algorithm from a general algorithm to a more specific one, by removing some functionalities, adapting the code for Android devices and optimizing its speed in exchange for marginally smaller face detection rate. Experimental results pointed out that we can significantly improve the speed of the algorithm by adapting it to the target device (we gained by face localization up to 13 ms with respect to the original algorithm), for the price of 18% reduction in face detection rate